|
| 1 | +""" |
| 2 | +Model Freezing in TorchScript |
| 3 | +============================= |
| 4 | +
|
| 5 | +In this tutorial, we introduce the syntax for *model freezing* in TorchScript. |
| 6 | +Freezing is the process of inlining Pytorch module parameters and attributes |
| 7 | +values into the TorchScript internal representation. Parameter and attribute |
| 8 | +values are treated as final values and they cannot be modified in the resulting |
| 9 | +Frozen module. |
| 10 | +
|
| 11 | +Basic Syntax |
| 12 | +------------ |
| 13 | +Model freezing can be invoked using API below: |
| 14 | +
|
| 15 | + ``torch.jit.freeze(mod : ScriptModule, names : str[]) -> SciptModule`` |
| 16 | +
|
| 17 | +Note the input module can either be the result of scripting or tracing. |
| 18 | +See https://pytorch.org/tutorials/beginner/Intro_to_TorchScript_tutorial.html |
| 19 | +
|
| 20 | +Next, we demonstrate how freezing works using an example: |
| 21 | +""" |
| 22 | + |
| 23 | +import torch, time |
| 24 | + |
| 25 | +class Net(torch.nn.Module): |
| 26 | + def __init__(self): |
| 27 | + super(Net, self).__init__() |
| 28 | + self.conv1 = torch.nn.Conv2d(1, 32, 3, 1) |
| 29 | + self.conv2 = torch.nn.Conv2d(32, 64, 3, 1) |
| 30 | + self.dropout1 = torch.nn.Dropout2d(0.25) |
| 31 | + self.dropout2 = torch.nn.Dropout2d(0.5) |
| 32 | + self.fc1 = torch.nn.Linear(9216, 128) |
| 33 | + self.fc2 = torch.nn.Linear(128, 10) |
| 34 | + |
| 35 | + def forward(self, x): |
| 36 | + x = self.conv1(x) |
| 37 | + x = torch.nn.functional.relu(x) |
| 38 | + x = self.conv2(x) |
| 39 | + x = torch.nn.functional.max_pool2d(x, 2) |
| 40 | + x = self.dropout1(x) |
| 41 | + x = torch.flatten(x, 1) |
| 42 | + x = self.fc1(x) |
| 43 | + x = torch.nn.functional.relu(x) |
| 44 | + x = self.dropout2(x) |
| 45 | + x = self.fc2(x) |
| 46 | + output = torch.nn.functional.log_softmax(x, dim=1) |
| 47 | + return output |
| 48 | + |
| 49 | + @torch.jit.export |
| 50 | + def version(self): |
| 51 | + return 1.0 |
| 52 | + |
| 53 | +net = torch.jit.script(Net()) |
| 54 | +fnet = torch.jit.freeze(net) |
| 55 | + |
| 56 | +print(net.conv1.weight.size()) |
| 57 | +print(net.conv1.bias) |
| 58 | + |
| 59 | +try: |
| 60 | + print(fnet.conv1.bias) |
| 61 | + # without exception handling, prints: |
| 62 | + # RuntimeError: __torch__.z.___torch_mangle_3.Net does not have a field |
| 63 | + # with name 'conv1' |
| 64 | +except RuntimeError: |
| 65 | + print("field 'conv1' is inlined. It does not exist in 'fnet'") |
| 66 | + |
| 67 | +try: |
| 68 | + fnet.version() |
| 69 | + # without exception handling, prints: |
| 70 | + # RuntimeError: __torch__.z.___torch_mangle_3.Net does not have a field |
| 71 | + # with name 'version' |
| 72 | +except RuntimeError: |
| 73 | + print("method 'version' is not deleted in fnet. Only 'forward' is preserved") |
| 74 | + |
| 75 | +fnet2 = torch.jit.freeze(net, ["version"]) |
| 76 | + |
| 77 | +print(fnet2.version()) |
| 78 | + |
| 79 | +B=1 |
| 80 | +warmup = 1 |
| 81 | +iter = 1000 |
| 82 | +input = torch.rand(B, 1,28, 28) |
| 83 | + |
| 84 | +start = time.time() |
| 85 | +for i in range(warmup): |
| 86 | + net(input) |
| 87 | +end = time.time() |
| 88 | +print("Scripted - Warm up time: {0:7.4f}".format(end-start), flush=True) |
| 89 | + |
| 90 | +start = time.time() |
| 91 | +for i in range(warmup): |
| 92 | + fnet(input) |
| 93 | +end = time.time() |
| 94 | +print("Frozen - Warm up time: {0:7.4f}".format(end-start), flush=True) |
| 95 | + |
| 96 | +start = time.time() |
| 97 | +for i in range(iter): |
| 98 | + input = torch.rand(B, 1,28, 28) |
| 99 | + net(input) |
| 100 | +end = time.time() |
| 101 | +print("Scripted - Inference: {0:5.2f}".format(end-start), flush=True) |
| 102 | + |
| 103 | +start = time.time() |
| 104 | +for i in range(iter): |
| 105 | + input = torch.rand(B, 1,28, 28) |
| 106 | + fnet2(input) |
| 107 | +end = time.time() |
| 108 | +print("Frozen - Inference time: {0:5.2f}".format(end-start), flush =True) |
| 109 | + |
| 110 | +############################################################### |
| 111 | +# On my machine, I measured the time: |
| 112 | +# |
| 113 | +# * Scripted - Warm up time: 0.0107 |
| 114 | +# * Frozen - Warm up time: 0.0048 |
| 115 | +# * Scripted - Inference: 1.35 |
| 116 | +# * Frozen - Inference time: 1.17 |
| 117 | + |
| 118 | +############################################################### |
| 119 | +# In our example, warm up time measures the first two runs. The frozen model |
| 120 | +# is 50% faster than the scripted model. On some more complex models, we |
| 121 | +# observed even higher speed up of warm up time. freezing achieves this speed up |
| 122 | +# because it is doing some the work TorchScript has to do when the first couple |
| 123 | +# runs are initiated. |
| 124 | +# |
| 125 | +# Inference time measures inference execution time after the model is warmed up. |
| 126 | +# Although we observed significant variation in execution time, the |
| 127 | +# frozen model is often about 15% faster than the scripted model. When input is larger, |
| 128 | +# we observe a smaller speed up because the execution is dominated by tensor operations. |
| 129 | + |
| 130 | +############################################################### |
| 131 | +# Conclusion |
| 132 | +# ----------- |
| 133 | +# In this tutorial, we learned about model freezing. Freezing is a useful technique to |
| 134 | +# optimize models for inference and it also can significantly reduce TorchScript warmup time. |
0 commit comments