CODEWITHSUNDEEP
pythonmachine-learningkeraspytorch
snehal desai
snehal desai

Reputation: 129

Converting From .pt model to .h5 model

I am using Google colab. I want to convert .pt model from my google drive to .h5 model. I follow link https://github.com/gmalivenko/pytorch2keras and https://www.programmersought.com/article/57938937172/ and install libraries and also write code as below:

%pip install pytorch2keras
%pip install onnx==1.8.1

import numpy as np
from numpy import random
from random import uniform
import torch
from torch.autograd import Variable
input_np = np.random.uniform(0, 1, (1, 10, 32, 32))
input_var = Variable(torch.FloatTensor(input_np))
model='/content/gdrive/MyDrive/model.pt'
pytorch_to_keras(model,input_var,input_shapes= [(10, 32, 32,)], verbose=True)

but it gives me error like:

WARNING:pytorch2keras:Custom shapes isn't supported now.
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-53-eef217a11c8a> in <module>()
      8 input_var = Variable(torch.FloatTensor(input_np))
      9 model='/content/gdrive/MyDrive/model.pt'
---> 10 pytorch_to_keras(model,input_np,input_shapes= [(10, 32, 32,)], verbose=True)
     11 
     12 

/usr/local/lib/python3.7/dist-packages/pytorch2keras/converter.py in pytorch_to_keras(model, args, input_shapes, change_ordering, verbose, name_policy, use_optimizer, do_constant_folding)
     51     args = tuple(args)
     52 
---> 53     dummy_output = model(*args)
     54 
     55     if isinstance(dummy_output, torch.autograd.Variable):

TypeError: 'str' object is not callable

Upvotes: 3

Views: 8034

Answers (4)

Richard Faucheron
Richard Faucheron

Reputation: 131

If you simply want to convert a .pt model state to a .h5, without worrying about Keras or ONNX compatibility, you can load the model state and then export each tensor recursively to a .h5 file:

import torch
import h5py

# Load your model state .pt
model_state_file = 'path/to/file.pt'
model_state = torch.load(model_state_file, map_location=torch.device('cpu'))

# Recursively save to .h5
def save_dict_h5py(h5py_file, group_name, dic, verbose=False):    
    for key, item in dic.items():
        if isinstance(item, (dict, collections.OrderedDict)):
            save_dict_h5py(h5py_file, group_name + '/' + str(key), item)
        else:
            if verbose: print(f'| writing {group_name + "/" + str(key)}')
            h5py_file[group_name + '/' + str(key)] = item

with h5py.File('path/to/file.h5', 'w') as f:
    save_dict_h5py(f, group_name='model_state', dic=model_state, verbose=True)

This function should print something like:

| writing model_state/fc1.0.weight
| writing model_state/fc1.0.bias
| writing model_state/unet_block.conv1.0.weight
| writing model_state/unet_block.conv1.0.bias

You can now load your .h5 file:

with h5py.File('path/to/file.h5', 'r') as f:
    print('fc1.0 bias shape:', f['model_state']['fc1.0.bias'])
    print('fc1.0 bias values:', f['model_state']['fc1.0.bias'][()])
    print('upconv8.0 weight shape:', f['model_state']['unet_block.upconv8.0.weight'])

# ------ Returns something like ------
# fc1.0 bias shape: <HDF5 dataset "fc1.0.bias": shape (16384,), type "<f4">
# fc1.0 bias values: [0.02804164 0.04839667 0.04619668 ... 0.04043121 0.03506938 0.05372864]
# upconv8.0 weight shape: <HDF5 dataset "unet_block.upconv8.0.weight": shape (512, 512, 9, 9), type "<f4">

Upvotes: 0

Carmine T. Guida
Carmine T. Guida

Reputation: 21

"model" should be a (loaded) pytorch model (not a string)

model = torch.load('your_model.py')
keras_model = pytorch_to_keras(model,input_var,input_shapes= [(10, 32, 32,)], verbose=True)

Upvotes: 0

Stanley
Stanley

Reputation: 839

Ah, the classic problem of PyTorch to Tensorflow. Many libraries have come and gone over the years, but I've found ONNX to work the most consistent. You could try something like this.

Specific to PyTorch is a Dynamic Computational Graph. A dynamic computational graph means that PyTorch models can dynamically adapt to different input sizes. You can specify which axes need dynamic sizing as such. Here is some minimal code to convert a CNN from PyTorch to ONNX.

import onnx
import torch

model = get_model()
model.eval()

# Test model on sample image size 
example_input = torch.randn((1, 3, img_size, img_size), requires_grad=True)
model(example_input)

# Set input and output names, include more names in the list if your model has more than 1 input/output
input_names = ["input0"]
output_names = ["output0"]

# Set dynamic axes (in this case, make the batch a dynamic dimension)
dynamic_axes = {'input0': {0: 'batch'}, 'output0': {0: 'batch'}}

# Export model with the above parameters
torch_out = torch.onnx.export(
    model, example_input, 'model.onnx', export_params=True, input_names=input_names, output_names=output_names, 
    dynamic_axes=dynamic_axes, operator_export_type=torch.onnx.OperatorExportTypes.ONNX
)

# Use ONNX checker to verify integrity of model
onnx_model = onnx.load("model.onnx")
onnx.checker.check_model(onnx_model)

One could also set the height and width of the model as a dynamic input size with

dynamic_axes['input0'][2] = 'height'
dynamic_axes['input0'][3] = 'width'

Next, we convert our ONNX model to a Tensorflow SavedModel.

from onnx_tf.backend import prepare
import onnx

onnx_model = onnx.load('model.onnx')
tf_model = prepare(onnx_model)
tf_model.export_graph('tf_model')

tf_model is now a Tensorflow SavedModel.

Upvotes: 4

franzmaliszt
franzmaliszt

Reputation: 41

Seems like you haven't imported the pytorch_to_keras method to begin with. Also, as the first parameter you might need to feed the actual model, not the model path. Give these a shot however, i am not familiar with this library.

Upvotes: 0

Related Questions