fixes from code review

minor changes to syntax, commenting, formatting, etc

adversarial_analysis.py

@@ -1,9 +1,7 @@
 import os
 import sys
 
-ROOT_DIR = os.getcwd()
-while 'DeepSparseCoding' in ROOT_DIR:
-    ROOT_DIR = os.path.dirname(ROOT_DIR)
+ROOT_DIR = os.path.dirname(os.getcwd())
 if ROOT_DIR not in sys.path: sys.path.append(ROOT_DIR)
 
 import numpy as np

models/base_model.py

@@ -124,7 +124,7 @@ def get_optimizer(self, optimizer_params, trainable_variables):
                 lr=optimizer_params.weight_lr,
                 weight_decay=optimizer_params.weight_decay)
         else:
-            assert False, ('optimizer name must be "sgd" or "adam", not %s'%(optimizer_name))
+            assert False, (f'optimizer name must be "sgd" or "adam", not {optimizer_name}')
         return optimizer
 
     def setup_optimizer(self):

models/mlp_model.py

@@ -1,7 +1,6 @@
 import numpy as np
 
 import torch
-#import torch.nn.functional as F
 
 from DeepSparseCoding.models.base_model import BaseModel
 from DeepSparseCoding.modules.mlp_module import MlpModule
@@ -15,7 +14,6 @@ def setup(self, params, logger=None):
     def get_total_loss(self, input_tuple):
         input_tensor, input_label = input_tuple
         pred = self.forward(input_tensor)
-        #return F.nll_loss(pred, input_label)
         self.loss_fn = torch.nn.CrossEntropyLoss()
         return self.loss_fn(pred, input_label)
 

modules/activations.py

@@ -11,7 +11,7 @@ def activation_picker(activation_function):
       return F.leaky_relu
     if activation_function == 'lca_threshold':
       return lca_threshold
-    assert False, ('Activation function ' + activation_function + ' is not supported.')
+    assert False, (f'Activation function {activation_function} is not supported.')
 
 def lca_threshold(u_in, thresh_type, rectify, sparse_threshold):
     u_zeros = torch.zeros_like(u_in)
@@ -38,5 +38,5 @@ def lca_threshold(u_in, thresh_type, rectify, sparse_threshold):
                     u_in,
                     u_zeros))
     else:
-        assert False, ('Parameter thresh_type must be "soft" or "hard", not '+thresh_type)
+        assert False, (f'Parameter thresh_type must be "soft" or "hard", not {thresh_type}')
     return a_out

modules/mlp_module.py

@@ -31,7 +31,6 @@ def preprocess_data(self, input_tensor):
     def forward(self, x):
         for dropout, act_func, layer in zip(self.dropout, self.act_funcs, self.layers):
             x = dropout(act_func(layer(x)))
-        #x = F.log_softmax(x, dim=1)
         return x
 
     def get_encodings(self, input_tensor):

params/base_params.py

@@ -6,6 +6,61 @@
 
 
 class BaseParams(object):
+    """
+    all models
+        batch_size [int] number of images in a training batch
+        data_dir [str] location of dataset folders
+        device [str] which device to run on
+        dtype [torch dtype] dtype for network variables
+        eps [float] small value to avoid division by zero
+        lib_root_dir [str] system location of this library directory
+        log_to_file [bool] if set, log to file, else log to stderr
+        model_name [str] name for model (can be anything)
+        model_type [str] type of model (must be among the list returned by utils/loaders.get_module_list())
+        num_epochs [int] how many epochs to use for training
+        num_pixels [int] total number of pixels in the input image
+        out_dir [str] base directory for all model outputs
+        optimizer [object] empty object with the following properties:
+            optimizer.name [str] which optimization algorithm to use
+                can be "sgd" (default) or "adam"
+            optimizer.lr_annealing_milestone_frac [list of floats] fraction of num_epochs to anneal learning rate
+            optimizer.lr_decay_rate [float] amount to anneal learning rate by
+        rand_seed [int] seed to be given to np.random.RandomState
+        rand_state [int] random state to be used for all random functions to allow for reproducible results
+        renormalize_weights [bool] if set, l2 normalize weights after each update
+        rescale_data_to_one [bool] if set, rescale input data to be between 0 and 1, per example
+        version [str] model version for output
+        shuffle_data [bool] if set, shuffle loader data before delivering batches
+        standardize_data [bool] if set, z-score data to have mean=0 and standard deviation=1 using numpy operators
+        train_logs_per_epoch [int or None] how often to send updates to the logfile
+        workspace_dir [str] system directory that is the parent to the primary repository directory
+
+    mlp
+        activation_functions [list of str]  strings correspond to activation functions for layers.
+            len must equal the len of layer_types
+            strings must be one of those listed in modules/activations.activation_picker()
+        dropout_rate [list of floats] specifies dropout probability of a value being set to zero or None per layer
+            len must be equal to the len of layer_types
+        layer_types [list of str] weight connectivity type, either "conv" or "fc"
+            len must be equal to the len of layer_channels - 1
+        layer_channels [list of int] number of outputs per layer, including the input layer
+
+    lca
+        dt [float] discrete global time constant for neuron dynamics
+            lca update rule is multiplied by dt/tau
+        num_latent [int] number of lca latent units
+        num_steps [int] number of lca inference steps to take
+        rectify_a [bool] if set, rectify the layer 1 neuron activity
+        sparse_mult [float] multiplyer placed in front of the sparsity loss term
+        tau [float] LCA time constant
+            lca update rule (step_size) is multiplied by dt/tau
+        thresh_type [str] specifying LCA threshold function; can be "hard" or "soft"
+
+    lca; mlp
+        weight_decay [float] multiplier to use on top of weight decay loss term
+        weight_lr [float] learning rate for all weight updates
+    """
+
     def __init__(self):
         self.set_params()
         self.compute_helper_params()

requirements.txt

@@ -1,4 +1,5 @@
 DeepSparseCoding
+foolbox>=3.1.1
 h5py>=2.8.0
 imageio>=2.4.1
 ipython>=7.1.1
@@ -11,7 +12,8 @@ scikit-image>=0.14.1
 scikit-learn>=0.20.0
 scipy>=1.1.0
 seaborn>=0.9.0
-tensorboard==2.2
-tensorflow-gpu==2.1
+tensorflow-gpu==1.15.2
+tensorflow-estimator==1.15.1
+tensorboard==1.15
+tensorflow-probability==0.8.0
 tensorflow-compression
-tensorflow-probability

tests/test_data_processing.py

@@ -185,7 +185,7 @@ def test_standardize(self):
                             places=num_tolerance_decimals,
                             msg=err_msg)
 
-    def rescale_data_to_one(self):
+    def test_rescale_data_to_one(self):
         num_tolerance_decimals = 7
         unflat_shape = [8, 4, 4, 3]
         flat_shape = [8, 4*4*3]

tests/test_models.py

@@ -20,7 +20,7 @@ def setUp(self):
 
     def test_model_loading(self):
         for model_type in self.model_list:
-            model_type = ''.join(model_type.split("_")[:-1]) # remove '_model' at the end
+            model_type = ''.join(model_type.split('_')[:-1]) # remove '_model' at the end
             model = loaders.load_model(model_type)
             params = loaders.load_params(self.test_params_file, key=model_type+'_params')
             train_loader, val_loader, test_loader, data_params = datasets.load_dataset(params)
@@ -31,7 +31,7 @@ def test_model_loading(self):
     ### TODO - more basic test to compute gradients per model###
     #def test_gradients(self):
     #    for model_type in self.model_list:
-    #        model_type = ''.join(model_type.split("_")[:-1]) # remove '_model' at the end
+    #        model_type = ''.join(model_type.split('_')[:-1]) # remove '_model' at the end
     #        model = loaders.load_model(model_type)
 
     def test_lca_ensemble_gradients(self):

tf1x/analyze_model.py

@@ -5,8 +5,8 @@
 import numpy as np
 import tensorflow as tf
 
-root_path = os.path.dirname(os.path.dirname(os.getcwd()))
-if root_path not in sys.path: sys.path.append(root_path)
+ROOT_DIR = os.path.dirname(os.path.dirname(os.getcwd()))
+if ROOT_DIR not in sys.path: sys.path.append(ROOT_DIR)
 
 from DeepSparseCoding.tf1x.utils.logger import Logger
 import DeepSparseCoding.tf1x.utils.data_processing as dp