Merge pull request #2 from deepvision-class/A3

A3

coutils/__init__.py

@@ -1,3 +1,5 @@
-from .import_colab import register_colab_notebooks
+from .import_colab import extract_drive_file_id, register_colab_notebooks
 from .utils import tensor_to_image, fix_random_seed, visualize_dataset
+from .eval import rel_error, compute_numeric_gradient
+from .solver import Solver
 from . import data

coutils/eval.py

@@ -0,0 +1,63 @@
+import torch
+
+@torch.no_grad()
+def rel_error(pred, true, eps=1e-8):
+  """
+  Returns relative error: (|pred| - |true|) / |true|
+
+  Inputs:
+  - pred: A torch tensor of the prediction of the true tensor
+  - true: A torch tensor of the true tensor
+  - eps: Small epsilon for numerical stability
+
+  Returns:
+  - Relative error between pred and true in floating point
+  """
+  return ((pred-true).abs() / true.abs().clamp(min=eps)).max().item()
+
+
+@torch.no_grad()
+def compute_numeric_gradient(f, x, dy=None, h=1e-5):
+  """
+  Compute a numeric gradient of the function f at the point x.
+  Inputs:
+  - f: The function for which to compute the numeric gradient
+  - x: A torch tensor giving the point at which to evaluate the gradient
+  - dy: (Optional) A torch tensor giving the upstream gradient dL/dy. If this
+    is not provided then initialize it to be all ones.
+  - h: (Optional) A Python float giving the step size to use 
+  
+  Returns:
+  - dx: A torch tensor giving the downstream gradient dL/dx
+  """
+  # First run the function unmodified
+  y = f(x)
+
+  # Initialize upstream gradient to all ones if not provided
+  if dy is None:
+    dy = torch.ones_like(y)
+
+  # Initialize downstream gradient to zeros
+  dx = torch.zeros_like(x)
+
+  # Get flattened views of everything
+  x_flat = x.contiguous().view(-1)
+  y_flat = y.contiguous().view(-1)
+  dx_flat = dx.contiguous().view(-1)
+  dy_flat = dy.contiguous().view(-1)
+  for i in range(dx_flat.shape[0]):
+    # Compute numeric derivative dy/dx[i]
+    orig = x_flat[i].item()
+    x_flat[i] = orig + h
+    yph = f(x).clone().view(-1)
+    x_flat[i] = orig - h
+    ymh = f(x).clone().view(-1)
+    x_flat[i] = orig
+    dy_dxi = (yph - ymh) / (2.0 * h) 
+
+    # Use chain rule to compute dL/dx[i] =  dL/dy . dy/dx[i]
+    dx_flat[i] = dy_flat.dot(dy_dxi).item()
+
+  dx = dx_flat.view(x.shape)
+  return dx
+

coutils/import_colab.py

@@ -11,6 +11,14 @@
 from oauth2client.client import GoogleCredentials
 
 
+def extract_drive_file_id(notebook_link):
+  prefix = 'https://colab.research.google.com/drive/'
+  if not notebook_link.startswith(prefix):
+    raise ValueError('notebook_link should start with "%s"' % prefix)
+  file_id = notebook_link[len(prefix):]
+  return file_id
+
+
 def register_colab_notebooks(id_map):
   """
   Register a set of Colab notebooks so they can be imported as Python modules.
@@ -30,6 +38,8 @@ def register_colab_notebooks(id_map):
   Inputs:
   - id_map: Dictionary mapping module names to Google Drive IDs
   """
+  # Clear any existing ColabNotebookFinder objects from the path
+  sys.meta_path = [x for x in sys.meta_path if not isinstance(x, ColabNotebookFinder)]
   sys.meta_path.append(ColabNotebookFinder(id_map))
 
 
@@ -70,7 +80,12 @@ def load_module(self, name):
         if cell.cell_type == 'code':
           itm = self.shell.input_transformer_manager
           code = itm.transform_cell(cell.source)
-          exec(code, mod.__dict__)
+          # exec import, class, or function cell only
+          if code.startswith('import ') or code.startswith('class ') or code.startswith('def ') or code.startswith('@'):
+            exec(code, mod.__dict__)
+    except Exception as e:
+      print(e)
+      print(code)
     finally:
       self.shell.user_ns = save_user_ns
     return mod
@@ -87,4 +102,3 @@ def find_module(self, fullname, path=None):
     loader = ColabLoader(self.id_map)
     return loader
 
-