Deliverable results for GPU training: README_GPU and supporting document in folder 1_gpu_training
Deliverable results for NPU training: README_NPU and supporting document in folder 2_npu_training
Deliverable results for NPU infernce: README_INFERENCE and supporting document in folder 3_npu_310_inference
| Issue | Code change |
|---|---|
| There seems to be conflicts if you use pywrap_tensorflow.NewCheckpointReader to load pretrained model and create a tf variable in your own model that has the same name with the one in the pretrained model but with different shapes. | Assign the variable created with a different name from the tensor loaded from the pretrained modoel |
| slicing tensor. Not support pythonic way of slicing tensor. Happens when running sess.run(tf.assign(v[:,:,:3,:],original_v)) where v is the layer in the model with shape (7,7,20,64) and original_v is the layer in the resnet50 with shape (7,7,3,64) | fix this issue by changing sess.run(tf.assign(v[:,:,:3,:],original_v) to sess.run(tf.assign(v[0:7,0:7,0:3,0:64],original_v). |
| tf.scatter_nd does not check indices as it does when using gpu training. | Remove input which exceeds the indice range. |