The structure of my resnet is different from built in PocketFlow，I how to change the structure? about pocketflow HOT 11 CLOSED

Hi, are you using checkpoint files produced by your own training code, instead of pre-trained models provided by us? This will cause the above error message.

If you do need to use your own model definition and pre-trained models, then you need to create your own ModelHelper class and a Python script to use it, similar to:

1. nets/resnet_at_cifar10.py (which defines a ModelHelper class)
2. nets/resnet_at_cifar10_run.py (which uses the above class)

from pocketflow.

I download models_resnet_56_at_cifar_10.tar.gz from https://api.ai.tencent.com/pocketflow/list.html, and decompress it in models.

error:
InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Assign requires shapes of both tensors to match. lhs shape= [32] rhs shape= [16]
[[Node: model/save/Assign_13 = Assign[T=DT_FLOAT, _class=["loc:@model/resnet_model/batch_normalization_11/gamma"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](model/resnet_model/batch_normalization_11/gamma, model/save/RestoreV2/_27)]]
[[Node: model/save/RestoreV2/_98 = _SendT=DT_FLOAT, client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_104_model/save/RestoreV2", _device="/job:localhost/replica:0/task:0/device:CPU:0"]]

from pocketflow.

you have to modify the layer num in resnet.py，the default layer num should be 50 or 101

from pocketflow.

Hi @as754770178
For nets/resnet_at_cifar10_run.py, the default number of layers is 20. Since you have downloaded the ResNet-56 model, you need to specify the number of layers with --resnet_size 56.

from pocketflow.

Thanks. I misunderstood the function of PocketFlow, I think the net defined in nets/resnet_at_cifar10_run.py is the student net. Actually, PocketFlow Pruning/Quantization the net defined in nets/resnet_at_cifar10_run.py as the student net? Is my idea correct?

from pocketflow.

I'm not sure whether I have understood your question.

In PocketFlow, the student network and teacher network (only exists if network distillation is enabled) share the same network architecture. The student network may have further restrictions introduced by pruning or quantization operations, while the teacher network is the full-precision uncompressed network. Does this resolve your question?

from pocketflow.

I want to confirm that the student only come from the pruning or quantization operations of the full-precision uncompressed network in the network distillation.

from pocketflow.

Yes, for all model compression methods in PocketFlow, the compressed network (or student network) only comes from the pruned / quantized version of a full-precision uncompressed network (or teacher network). This is irrelevant to network distillation, which only adds a distillation loss term in the training of student network.

from pocketflow.

ok, thanks

from pocketflow.

I define my net, but the name of variable is Prefixed of 'model', such as 'model/resnet_v1_110/block1/unit_1/bottleneck2_v1/conv1/BatchNorm/beta ', but it should is 'resnet_v1_110/block1/unit_1/bottleneck2_v1/conv1/BatchNorm/beta '.

`NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key model/resnet_v1_110/block1/unit_1/bottleneck2_v1/conv1/BatchNorm/beta not found in checkpoint
[[Node: model/save/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, ..., DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_model/save/Const_0_0, model/save/RestoreV2/tensor_names, model/save/RestoreV2/shape_and_slices)]]
[[Node: model/save/RestoreV2/_1059 = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_1064_model/save/RestoreV2", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]]`

from pocketflow.

It seems this issue has been resolved in #27. Closing. Reopen it if there are any further questions.

from pocketflow.