diff --git a/models/loss.py b/models/loss.py
new file mode 100644
index 0000000000000000000000000000000000000000..cea2ecbd0857773e025b133973113f4b1110026e
--- /dev/null
+++ b/models/loss.py
@@ -0,0 +1,30 @@
+import torch
+
+def get_memory_loss(memory_att):
+    """The memory attribute should be with size [batch_size, memory_dim, reduced_time_dim, f_h, f_w]
+    loss = \sum_{t=1}^{reduced_time_dim} (-mem) * (mem + 1e-12).log() 
+    averaged on each pixel and each batch
+    2. average over batch_size * fh * fw
+    """
+    s = memory_att.shape
+    memory_att = (-memory_att) * (memory_att + 1e-12).log()  # [batch_size, memory_dim, time, fh, fw]
+    memory_att = memory_att.sum() / (s[0] * s[-2] * s[-1]) 
+    return memory_att 
+    
+    
+def get_unormalized_data(x_input, x_recons, mean, std):
+    x_input = x_input.mul(std).add(mean)
+    x_recons = x_recons.mul(std).add(mean)
+    return x_input, x_recons
+
+
+def get_reconstruction_loss(x_input, x_recons, mean=0.5, std=0.5):
+    """Calculates the reconstruction loss between x_input and x_recons
+    x_input: [batch_size, ch, time, imh, imw]
+    x_recons: [batch_size, ch, time, imh, imw]
+    """
+    batch_size, ch, time_dimension, imh, imw = x_input.shape
+    x_input, x_recons = get_unormalized_data(x_input, x_recons, mean, std)
+    recons_loss = (x_input - x_recons) ** 2
+    recons_loss = recons_loss.sum().sqrt()/(batch_size * imh * imw)
+    return recons_loss
\ No newline at end of file