fastmri variant isolating

algorithmic_efficiency/workloads/fastmri/fastmri_jax/models.py

@@ -12,6 +12,7 @@
 Data:
 github.com/facebookresearch/fastMRI/tree/main/fastmri/data
 """
+import functools
 from typing import Optional
 
 import flax.linen as nn
@@ -65,6 +66,8 @@ class UNet(nn.Module):
   channels: int = 32
   num_pool_layers: int = 4
   dropout_rate: Optional[float] = 0.0  # If None, defaults to 0.0.
+  use_tanh: bool = False
+  use_layer_norm: bool = False
 
   @nn.compact
   def __call__(self, x, train=True):
@@ -134,6 +137,8 @@ class ConvBlock(nn.Module):
   """
   out_channels: int
   dropout_rate: float
+  use_tanh: bool
+  use_layer_norm: bool
 
   @nn.compact
   def __call__(self, x, train=True):
@@ -149,29 +154,35 @@ def __call__(self, x, train=True):
         features=self.out_channels,
         kernel_size=(3, 3),
         strides=(1, 1),
-        use_bias=False)(
-            x)
-    # InstanceNorm2d was run with no learnable params in reference code
-    # so this is a simple normalization along channels
-    x = _simple_instance_norm2d(x, (1, 2))
-    x = jax.nn.leaky_relu(x, negative_slope=0.2)
+        use_bias=False)(x)
+    if self.use_layer_norm:
+      x = nn.LayerNorm()(x)
+    else:
+      # DO NOT SUBMIT check that this comment edit is correct
+      # InstanceNorm2d was run with no learnable params in reference code
+      # so this is a simple normalization along spatial dims.
+      x = _instance_norm2d(x, (1, 2))
+    if self.use_tanh:
+      activation_fn = nn.tanh
+    else:
+      activation_fn = functools.partial(jax.nn.leaky_relu, negative_slope=0.2)
+    x = activation_fn(x)
     # Ref code uses dropout2d which applies the same mask for the entire channel
     # Replicated by using broadcast dims to have the same filter on HW
     x = nn.Dropout(
-        self.dropout_rate, broadcast_dims=(1, 2), deterministic=not train)(
-            x)
+        self.dropout_rate, broadcast_dims=(1, 2), deterministic=not train)(x)
     x = nn.Conv(
         features=self.out_channels,
         kernel_size=(3, 3),
         strides=(1, 1),
-        use_bias=False)(
-            x)
-    x = _simple_instance_norm2d(x, (1, 2))
-    x = jax.nn.leaky_relu(x, negative_slope=0.2)
+        use_bias=False)(x)
+    if self.use_layer_norm:
+      x = nn.LayerNorm()(x)
+    else:
+      x = simple_instance_norm2d(x, (1, 2))
+    x = activation_fn(x)
     x = nn.Dropout(
-        self.dropout_rate, broadcast_dims=(1, 2), deterministic=not train)(
-            x)
-
+        self.dropout_rate, broadcast_dims=(1, 2), deterministic=not train)(x)
     return x
 
 
@@ -180,6 +191,8 @@ class TransposeConvBlock(nn.Module):
   out_channels: Number of channels in the output.
   """
   out_channels: int
+  use_tanh: bool
+  use_layer_norm: bool
 
   @nn.compact
   def __call__(self, x):
@@ -192,7 +205,13 @@ def __call__(self, x):
     x = nn.ConvTranspose(
         self.out_channels, kernel_size=(2, 2), strides=(2, 2), use_bias=False)(
             x)
-    x = _simple_instance_norm2d(x, (1, 2))
-    x = jax.nn.leaky_relu(x, negative_slope=0.2)
-
+    if self.use_layer_norm:
+      x = nn.LayerNorm()(x)
+    else:
+      x = simple_instance_norm2d(x, (1, 2))
+    if self.use_tanh:
+      activation_fn = nn.tanh
+    else:
+      activation_fn = functools.partial(jax.nn.leaky_relu, negative_slope=0.2)
+    x = activation_fn(x)
     return x