add tools for converting model format (from pytorch to niutensor)

tools/pack_model.py

+import argparse
+from struct import pack
+
+import torch
+
+parser = argparse.ArgumentParser(description='Pack Pytorch model to NiuTensor')
+parser.add_argument('-src', help='pytorch model', type=str, default='model.pt')
+parser.add_argument('-tgt', help='niutensor model', type=str, default='model.bin')
+args = parser.parse_args()
+
+model = torch.load(args.src, map_location='cpu')
+
+model = model['model']
+
+def get_model_parameters(m):
+    '''
+    get flattend transformer model parameters
+    '''
+    p = []
+    w = None
+    for k in m:
+        if 'embed_tokens.weight'  in k:
+            w = m[k]
+        elif m[k].numel() != 1:
+            # p.append(m[k])
+            if 'weight' in k:
+                # weights for qkv
+                if 'in_proj' in k:
+                    dim = m[k].shape[0] // 3
+                    p.append((m[k][:dim, :]).t())
+                    p.append((m[k][dim:dim*2, :]).t())
+                    p.append((m[k][dim*2:, :]).t())
+                else:
+                    if 'norm' in k:
+                        p.append(m[k])
+                    else:
+                        p.append(m[k].t())
+            else:
+                p.append(m[k])
+    
+    # encoder embedding weight
+    p.append(w)
+    # decoder embedding weight
+    p.append(w)
+    # output weight
+    p.append(w)
+    return p
+
+with torch.no_grad():
+    params = get_model_parameters(model)
+
+    params_number = pack("Q", len(params))
+    params_size = pack("Q" * len(params), *[p.numel() for p in params])
+
+    print('total params: ', len(params))
+    print('total params size: ', sum([p.numel() for p in params]))
+    with open(args.tgt+".name.txt", "w") as name_list:
+        for p in model:
+            name_list.write("{}\t{}\n".format(p, model[p].shape))
+            
+    with open(args.tgt+".bin", 'wb') as tgt:
+        # part 1: number of parameters
+        # tgt.write(params_number)
+
+        # part 2: offsets of parameters
+        # tgt.write(params_size)
+
+        # part 3: values of parameters
+        for p in params:
+            values = pack("f" * p.numel(), *(p.contiguous().view(-1).cpu().tolist()))
+            tgt.write(values)
\ No newline at end of file