small bug fixes

README.md

@@ -32,7 +32,7 @@ However, Poisson likelihood is given by the formula:
 If taking the logarithm (log-likelihood), then this becomes:
 ```l(y) = -log(y!) + y*log(yhat) - yhat```
 
-Since `log(0!) = 0`, and the sum of predictions for all combinations of users and items can be quickly calculated by `sum yhat = sum_{i,j} <U_i, V_j> = <sum_i U_i, sum_j V_j>` (`U` and `V` are non-negative matrices), it means the model doesn't ever need to make calculations on values that are equal to zero - simply not adding them to calculations would implicitly assume that they are zero.
+Since `log(0!) = 0`, and the sum of predictions for all combinations of users and items can be quickly calculated by `sum yhat = sum_{i,j} <U_i, V_j> = <sum_i U_i, sum_j V_j>` (since `U` and `V` are non-negative matrices), it means the model doesn't ever need to make calculations on values that are equal to zero - simply not adding them to calculations would implicitly assume that they are zero.
 
 Moreover, negative Poisson log-likelihood is a more appropriate loss for count data than squared loss, which tends to produce not-so-good results when the values to predict follow an exponential rather than a normal distribution.
 
@@ -139,6 +139,21 @@ For a more detailed example, see the IPython notebook [recommending songs with E
 
 This package contains only functionality related to fitting this model. For general evaluation metrics for recommendations on implicit data see other packages such as [lightFM](https://github.com/lyst/lightfm).
 
+## Saving model with pickle
+
+Using pickle to save an `HPF` object might fail due to problems with lambda functions. The following solves it:
+
+```python
+import pickle
+from hpfrec import HPF
+
+h = HPF()
+h.step_size = None
+pickle.dump(h, open("HPF_obj.p", "wb"))
+```
+
+(Be aware though that afterwards it won't be possible to use `partial_fit` or `add_user` with updates to item parameters.)
+
 ## Documentation
 
 Documentation is available at readthedocs: [http://hpfrec.readthedocs.io](http://hpfrec.readthedocs.io/en/latest/)

hpfrec/__init__.py

@@ -343,7 +343,7 @@ def fit(self, counts_df, val_set=None):
 
         ## after terminating optimization
         if self.keep_data:
-            if self.users_per_batch != 0:
+            if self.users_per_batch == 0:
                 self._store_metadata()
             else:
                 self._st_ix_user = self._st_ix_user[:-1]

hpfrec/cython_loops.pyx

@@ -303,9 +303,9 @@ def fit_hpf(float a, float a_prime, float b_prime,
 		np.savetxt(os.path.join(save_folder, "Gamma_rte.csv"), Gamma_rte, fmt="%.10f", delimiter=',')
 		np.savetxt(os.path.join(save_folder, "Lambda_shp.csv"), Lambda_shp, fmt="%.10f", delimiter=',')
 		np.savetxt(os.path.join(save_folder, "Lambda_rte.csv"), Lambda_rte, fmt="%.10f", delimiter=',')
-		np.savetxt(os.path.join(save_folder, "kappa_rte.csv"), k_rte.reshape(-1,1), fmt="%.10f", delimiter=',')
-		np.savetxt(os.path.join(save_folder, "tau_rte.csv"), t_rte.reshape(-1,1), fmt="%.10f", delimiter=',')
-		np.savetxt(os.path.join(save_folder, "Phi.csv"), phi/Y, fmt="%.10f", delimiter=',')
+		np.savetxt(os.path.join(save_folder, "kappa_rte.csv"), k_rte.reshape((-1,1)), fmt="%.10f", delimiter=',')
+		np.savetxt(os.path.join(save_folder, "tau_rte.csv"), t_rte.reshape((-1,1)), fmt="%.10f", delimiter=',')
+		np.savetxt(os.path.join(save_folder, "Phi.csv"), phi/Y.reshape((-1,1)), fmt="%.10f", delimiter=',')
 
 	## returning objects as needed
 	if keep_all_objs:
@@ -406,7 +406,7 @@ def calc_user_factors(float a, float a_prime, float b_prime,
 		Theta_prev = Theta.copy()
 
 	if return_all:
-		return Gamma_shp, Gamma_rte, phi/Y.reshape(-1, 1)
+		return Gamma_shp, Gamma_rte, phi/Y.reshape((-1, 1))
 	else:
 		return None
 

setup.py

@@ -27,7 +27,7 @@
    'scipy',
    'cython'
 ],
-  version = '0.2.1.1',
+  version = '0.2.1.2',
   description = 'Hierarchical Poisson matrix factorization for recommender systems',
   author = 'David Cortes',
   author_email = 'david.cortes.rivera@gmail.com',