fixed issues reported by Leo

Author: wangleiphy

.gitignore

@@ -1,3 +1,4 @@
 *~
 __pycache__
 .DS_Store
+data

ising.py 1_ising_TRG/ising.py

@@ -1,7 +1,9 @@
+import sys
+sys.path.insert(0, '../')
 import torch
 import numpy as np
 
-from rg import TRG
+from tensornets import TRG
 
 def build_tensor(K):
     lam = [torch.cosh(K)*2, torch.sinh(K)*2]
@@ -20,7 +22,7 @@ def build_tensor(K):
 if __name__=="__main__":
     import argparse
     parser = argparse.ArgumentParser(description='')
-    parser.add_argument("-chi", type=int, default=24, help="chi")
+    parser.add_argument("-chi", type=int, default=16, help="chi")
     parser.add_argument("-Niter", type=int, default=20, help="Niter")
     parser.add_argument("-use_checkpoint", action='store_true', help="use checkpoint")
     parser.add_argument("-float32", action='store_true', help="use float32")

args.py 2_variational_iPEPS/args.py

@@ -1,6 +1,9 @@
+import sys
+sys.path.insert(0, '../')
+
 import torch 
 import time
-from rg import CTMRG
+from tensornets import CTMRG
 from measure import get_obs
 from utils import symmetrize
 from args import args

ipeps.py 2_variational_iPEPS/ipeps.py

@@ -38,7 +38,7 @@ def save_checkpoint(checkpoint_path, model, optimizer):
              'optimizer' : optimizer.state_dict()}
     #print(model.state_dict().keys())
     torch.save(state, checkpoint_path)
-    print('model saved to %s' % checkpoint_path)
+    #print('model saved to %s' % checkpoint_path)
 
 def load_checkpoint(checkpoint_path, args, model):
     print( 'load old model from %s ' % checkpoint_path )

measure.py 2_variational_iPEPS/measure.py

@@ -1,6 +1,7 @@
 '''
 Variational PEPS with automatic differentiation and GPU support
 '''
+
 import io
 import torch
 import numpy as np
@@ -68,18 +69,19 @@
         Mpx = kron(sx, id2)
         Mpy = kron(sy, id2)
         Mpz = kron(sz, id2)
-        print (H)
     else:
         print ('what model???')
         sys.exit(1)
 
+    print ('Hamiltonian:\n', H)
+
     def closure():
         optimizer.zero_grad()
         start = time.time()
         loss, Mx, My, Mz = model.forward(H, Mpx, Mpy, Mpz, args.chi)
         forward = time.time()
         loss.backward()
-        print (model.A.norm().item(), model.A.grad.norm().item(), loss.item(), Mx.item(), My.item(), Mz.item(), torch.sqrt(Mx**2+My**2+Mz**2).item(), forward-start, time.time()-forward)
+        #print (model.A.norm().item(), model.A.grad.norm().item(), loss.item(), Mx.item(), My.item(), Mz.item(), torch.sqrt(Mx**2+My**2+Mz**2).item(), forward-start, time.time()-forward)
         return loss
 
     with io.open(key+'.log', 'a', buffering=1, newline='\n') as logfile:
@@ -91,6 +93,6 @@ def closure():
             with torch.no_grad():
                 En, Mx, My, Mz = model.forward(H, Mpx, Mpy, Mpz, args.chi if args.chi_obs is None else args.chi_obs)
                 Mg = torch.sqrt(Mx**2+My**2+Mz**2)
-                message = ('{} ' + 5*'{:.16f} ').format(epoch, En, Mx, My, Mz, Mg)
+                message = ('{} ' + 5*'{:.5f} ').format(epoch, En, Mx, My, Mz, Mg)
                 print ('epoch, En, Mx, My, Mz, Mg', message)
                 logfile.write(message + u'\n')

utils.py 2_variational_iPEPS/utils.py

@@ -12,6 +12,7 @@ Run this to compute the energy and specific heat of a 2D classical Ising model b
 
 
 ```bash
+$ cd 1_ising_TRG
 $ python ising.py 
 ```
 
@@ -23,20 +24,15 @@ Run this to optimize an iPEPS wavefuntion for 2D quantum Heisenberg model. Here,
 
 
 ```bash
-$ python main.py -D 3 -chi 30 
+$ cd 2_variational_iPEPS
+$ python variational.py -D 3 -chi 30 
 ```
 
-In case of a question, can type `python main.py -h`. It is also possible to supply your own Hamiltonian and measure other physical observables of interests. 
+In case of a question, you can type `python variational.py -h`. To make use GPU, you can add `-cuda <GPUID>`.  It is also possible to supply Hamiltonian of your own interests. 
 
 ### What is under the hood ?
 
-Reverse mode AD computes gradient accurately and efficiently for you! Check the codes in [adlib](https://github.com/wangleiphy/tensorgrad/tree/master/rg/adlib) for backward functions which propagate gradients through tensor network contractions.  
-
-
-
-### Explore more
-
-See [link]() for a Julia implementation
+Reverse mode AD computes gradient accurately and efficiently for you! Check the codes in [adlib](https://github.com/wangleiphy/tensorgrad/tree/master/tensornets/adlib) for backward functions which propagate gradients through tensor network contractions.  
 
 
 

main.py 2_variational_iPEPS/variational.py

@@ -83,7 +83,7 @@ def CTMRG(T, chi, max_iter, use_checkpoint=False):
         if (diff < threshold):
             break
         sold = s
-    print ('ctmrg iterations, diff, error', n, diff, truncation_error/n)
+    #print ('ctmrg converged at iterations %d to %.5e, truncation error: %.5f'%(n, diff, truncation_error/n))
 
     return C, E